How does Mixup & CutMix Augmentations work?

[kuri54]

Hi!(But it's already night in Japan)

I was wondering if you could tell me about Mixup & CutMix Augmentations?
What I want to know is about the behavior of 'switch_prob' in mixup_args.

For example, if I set mixup_args as follows, how does it work?

mixup_args = {
    'mixup_alpha': 0.3,
    'cutmix_alpha': 0.3,
    'cutmix_minmax': None,
    'prob': 1.0,
    'switch_prob': 0,
    'mode': 'elem',
    'label_smoothing': 0,
    'num_classes': 1000}

What happens if I have both mixup and cutmix enabled, but 'switch_prob' is 0?

From reading the timmdocs, I'm thinking that only mixup is enabled without switching to cutmix, but what do you think?

[rwightman]

@kuri54 as per code comments https://github.com/rwightman/pytorch-image-models/blob/master/timm/data/mixup.py#L98

[kuri54]

Thanks for the prompt reply!

Thank you for providing the code.
I checked and I guess cutmix doesn't work if switch_prob is 0.
(I've checked that use_cutmix is all False.)

This confirms that I will have to redo some experiments later in the week.
However, I think it's a step in the right direction.

Thank you very much!